AI for Solar Companies: Where It Pays From First Text to PTO

Solar runs on two clocks. The sale is decided in minutes and days. A home owner fills in a form. A shared lead goes to several installers. The first to reply with good questions often gets the consult.

The delivery takes weeks and months. First come the survey and the permit. Then the install and the inspection. Then comes interconnection, the hookup to the grid. Last is permission to operate (PTO). That is the utility's OK to switch the system on.

AI pays on both clocks. At the front, it answers each lead in minutes. It qualifies by text and books the consult. It sends the large or tricky jobs to a person. At the back, it keeps the buyer in the loop. It chases the papers. That is where signed deals quietly cancel.

All of it sits inside the consent rules. And the AI can never promise a saving or a tax credit.

What changed for solar in 2026?

Two things make AI worth more this year. One thing makes it riskier.

The US home clean energy credit (Section 25D) is gone for home owners. The IRS says it is not allowed for costs paid after 31 December 2025. A cost counts as paid when the install is done.

Leasing firms can still claim their own credit for longer. That pushes the market toward leases and power purchase agreements (PPAs). With a PPA, the owner buys the power. They don't buy the panels.

Home solar put in 995 MW in the second quarter of 2026. That was down 12% on a year before. The figures are from SEIA and Wood Mackenzie. Solar Power World reported them.

Meanwhile, the cost to win each buyer is rising. Wood Mackenzie expects it to jump 40% this year in home solar. It goes from a five-year low of $0.60 per watt to $0.84. Two things drive it. One is the shift to third-party ownership. The other is costly lead buys. Trade prices put a shared lead at roughly $25 to $100. Exclusive leads run $100 to $250. At that price, the first minutes after a lead comes in count for more.

The risk: talk about credits and rebates is now a minefield. A sales script written in 2024 is wrong in 2026. An AI that learned from it will repeat it.

Where does AI pay in a solar company?

Row 3 is the one most "AI for solar" lists miss. It also has the clearest proof behind it.

Two clocks: the sale and the delivery

The figure at the top lays both clocks on one line. It runs from the first minute to PTO. The left half is the sale. The right half is the delivery, and it is long.

NREL is a US national energy lab. It studied deals that cancel. It used data from four large installers (2017 to 2019). Here is what it found.

• 51% of contracts did not end in a built system. The rate at the middle installer was lower, at 33%.
• Deals that cancel added about $0.20 per watt to soft costs. Those are the costs that are not hardware.
• Direct spend to win each buyer was around $1,000 per project, on average. That was up to the signing.
• 73% of cancellations came early. They came before the permit, or the approval to build, was filed.

That last number puts the leak in the weeks after signing. The buyer hears nothing. Then they start to doubt.

That is a messaging problem. And messaging is what AI does well. It can send a plain weekly update. "Your permit went in Tuesday. The city tends to take this long. Here is what happens next." It can chase the missing power bill. It can answer "is this still happening?" at 9 p.m. It reads status from your project system. It does not guess.

Speed to lead, inside the consent rules

The Harvard Business Review audited 2,241 firms. Some tried to reach a lead within an hour. They were nearly seven times as likely to qualify it as firms that waited an hour longer. In solar, a shared lead goes to several installers at once. So an hour is a long time.

The consent rules come first. The TCPA is the US law on sales calls and texts. It limits calls made with an "artificial or prerecorded voice." In February 2024, the FCC ruled on that. It said AI voices count.

So an AI that calls a lead needs prior express consent. That is a clear yes, given before the call. For telemarketing, it needs prior express written consent. That is a signed yes, in writing.

There was a one-to-one consent rule for lead sellers, too. The FCC set it in December 2023. In January 2025, a US appeals court threw it out. That was the Eleventh Circuit. The FCC took it off the books in July 2025. But consent still has to cover your firm. It has to cover the kind of contact you make, too.

Solar sales calls are a top target for the law. Take a 2023 FTC sweep. One solar outfit made tens of millions of calls, the FTC said. The numbers were on the Do Not Call list. The FTC said it lied about ties to the power companies, too. And it said the outfit overstated savings. The fine was $13.8 million, partly suspended.

So the first step in any solar AI is a consent check. Run it per lead and per channel. Store the proof. State rules stack on top. Take Florida. It bars sales calls before 8 a.m. and after 8 p.m. It caps them at three in 24 hours on the same subject, too.

Route by deal size, not by who picks up

Many solar firms sell two things at once. One is a small, standard product. It could almost sell itself. The other is larger or commercial jobs. Those need an expert. Say every lead goes to the same closer. Then the expert spends the week on the small ones.

We designed a system for a solar firm with just that split. An AI qualifies each new lead by text. It works inside guardrails that rule out promises on price. Then it tags the lead as small or large.

Small prospects get an AI phone agent. It answers questions and handles objections. Then it sends a secure checkout link. The link is for a low-priced starter offer, on a page the payment firm hosts. Larger prospects are booked straight onto a person's calendar. The full context goes with them. Reminders and no-show follow-ups run on their own.

We estimate a system like this is worth $100K to $320K a year. That is for a solar firm of that shape. Two things drive that range. First, it takes the founder out of small-deal calls. That frees an estimated 15 to 25 hours a week. Second, no-show follow-up wins back pipeline that would be lost. We put that at 15% to 30%. This is an estimate from the design, not a result.

What the AI must never say

Savings and credits are where solar sales get sued. The Treasury, the CFPB and the FTC put out a joint warning. Some solar sellers deceived people on costs and savings, it said. They misled people on loans and tax credits, too. In May 2026, Texas's attorney general sued a solar firm. The suit says it misled people on who could get the tax credit. It says the firm misled them on savings, too.

An AI repeats what it is given, at scale. So give it hard limits.

• No savings figure, ever. Savings depend on the roof, the usage and the rate plan. The consultant shows a design with the buyer's own numbers.
• No tax-credit promises. The homeowner credit is gone for installs completed after 2025. And the math for leases and PPAs is not the same.
• No claim of ties to the utility. No "government program." No loan terms.
• Money, contracts and cancelling all go to a person.

A chatbot's words are the firm's words. In Moffatt v. Air Canada, a tribunal held the airline to what its chatbot said. The same logic will apply to a solar bot that quotes a saving.

After the contract: permits, interconnection and PTO

California shows how much paper the delivery half carries. The CPUC sets the rules for power firms there. It says solar buyers must get the state's disclosure papers. They get them before they sign. Then those papers go to the utility. If they don't, the grid hookup form is not complete. That form also needs a signed Solar Consumer Protection Guide.

An AI can check each packet for missing pieces before it goes in. It can chase the buyer to sign. And it can tell them what each step is for, in plain words.

Design and proposal software is mature, so buy it. The build that pays is the glue between your systems. It turns status from your project system into updates. It chases papers and answers questions. And it flags stuck jobs to a person. That happens before the buyer gives up.

What does it cost to run?

Our build library is the record we keep of each build. Take the system above, which screens and routes leads. The library puts its tools at $250 to $1,300 a month. That covers the CRM and a voice agent platform. The platform bills by call volume. It covers phone numbers and texting, plus payment fees. Voice is the part that moves. Text and status updates are cheap next to it.

The fixed cost is ownership. Someone has to keep the script up to date on credits and rules. They read a sample of chats each week. And they keep the consent records clean. In 2026, that person is worth more than any model upgrade.

What goes wrong

• Calls without the right consent. An AI voice is an artificial voice under the TCPA. Check consent before you dial.
• The 2024 script in a 2026 market. The AI offers a 30% credit to a cash buyer. Audit every claim it can make.
• Made-up savings. A cheerful number with no roof data behind it is a legal risk.
• A payment link to the wrong buyer. Self-serve checkout suits a standard, low-priced product. A full home system needs a person and a design. It needs the required disclosures, too.
• Silence after signing. No updates come for six weeks. The buyer cancels before the permit is filed.
• Status guessed, not read. The AI says "install next week," because that is the norm. It must read the project system.

Questions to ask a vendor

1. How does the system check consent before each automated call or text? Where is the proof stored?
2. Which claims is the AI blocked from making? Think savings, credits, loans and ties to others. Who updates that list?
3. What happens when a homeowner asks about tax credits?
4. How are leads routed between self-serve, a consultant and a specialist?
5. Does the AI read project status from our system, or does it guess?
6. How are quiet hours and call caps enforced? Is that by state and time zone?
7. Whose accounts do the phone numbers, the CRM and the bills sit on?

When not to buy AI

Too few leads. Take a full stack for screening and voice. Our build library's line for it is about 50 new leads a month. Below that, a simple automated text wins. So does a fast human.

Only high-ticket sales that need a consult. Does every deal need a design and a person? Then drop the AI phone agent. Build screening and booking only.

No lead flow yet. This system routes and converts leads. It does not create them.

Messy project data. Do permit and install status live in someone's inbox? Then fix the project system first. After that, AI can report from it.